Detection of Cardiac Infarction in MRI C-SENC Images

8/8/2019 Detection of Cardiac Infarction in MRI C-SENC Images

http://slidepdf.com/reader/full/detection-of-cardiac-infarction-in-mri-c-senc-images 1/5

Universal Journal of Computer Science and Engineering Technology

1 (1), 36-40, Oct. 2010.

© 2010 UniCSE, ISSN: 2219-2158

36

Corresponding Author: Ahmad O. Algohary, Senior Biomed. S/W Eng. Diagnosoft Inc, Cairo Intl Office, Egypt.

Detection of Cardiac Infarction in MRI C-SENC

Images

Ahmad O. Algohary

Senior Biomed S/W Eng

Diagnosoft Inc

Cairo Intl Office, Egypt

ahmad.algohary@diagno

soft.com

Ahmed M. El-Bialy

Systems and Biomedical

Eng. Dept., Faculty of

Engineering

Cairo Univ., Giza, Egypt

[email protected]

Ahmed H. Kandil

Systems and Biomedical

Eng. Dept., Faculty of

Engineering

Cairo Univ., Giza, Egypt

[email protected]

Nael F. Osman

Radiology Department,

School of Medicine

Johns Hopkins

University, Baltimore,

North Carolina, USA

Abstract — Composite Strain Encoding (C-SENC) is an Magnetic

Resonance Imaging (MRI) technique for acquiring simultaneous

viability and functional and images of the heart. It combines two

imaging techniques, Delayed Enhancement (DE) and Strain

Encoding (SENC). In this work, a novel multi-stage method is

proposed to identify ventricular infarction in the functional and

viability images provided by C-SENC MRI. The proposed

method is based on sequential application of Otsu’s thresholding,

morphological opening, square boundary tracing and the

subtractive clustering algorithm. This method is tested on images

of ten patients with and without myocardial infarction (MI). The

resulting clustered images are compared with those marked up

by expert cardiologists who assisted in validating results coming

from the proposed method. Infarcted tissues are correctly

identified using the proposed method with high levels of

sensitivity and specificity.

Keywords- Infarction; C-SENC MRI; Delayed Enhancement;

styling; insert (key words)

I. INTRODUCTION

CCURATE identification of myocardial infarction (MI) is

considered critical for therapeutical decision-making [1].

Functional images are useful in determining the contractility

patterns in the affected regions [1]. On the other hand,

viability images are used to differentiate between viable and

nonviable tissues [3].

By combining the information of functionality and viability,

myocardial tissue can be classified into three identifiable

types: (i) normally contracting tissue, which represents normal

myocardium; (ii) non-contracting yet viable tissue, which

represents hibernating myocardium; and (iii) non-viable tissue,

which represents infarcted myocardium [4].

Previously, a method was proposed to identify different

heart tissues from MRI C-SENC images using an

unsupervised multi-stage clustering technique [5]. The method

was based on sequential application of the Fuzzy C-means

(FCM) and iterative self-organizing data (ISODATA)

clustering algorithms. In a more recent work, Bayesian

classifier was proposed to identify the background region, then

the filtered tissue regions were classified into the different

tissue types using FCM algorithm [6].

In this work, a novel, automatic, multi-stage method isproposed to identify different heart tissues from tuned imagesprovided by Cardiac Magnetic Resonance (CMR) CompositeStrain Encoding (C-SENC) images of transverse sections of the

left ventricle (LV), and to identify infarcted myocardial tissues.This method is based on the application of Otsu’s thresholding technique, morphological opening, square boundary tracingand the subtractive clustering algorithm. Numericalsimulations, real CMR images of patients and expertcardiologists’ markings were used to validate the proposed method, which showed excellent results with respect tosensitivity and specificity.

II. THEORY

A. C-SENC

Recently, the Composite Strain Encoding (C-SENC) MRI

technique has been introduced for simultaneous cardiac

functional and viability imaging in a single short

breathhold [1]. No additional time, when compared with

standard Delayed Enhancement (DE) viability imaging, is

required for acquiring the additional functional images. This

technique results in three images: no-tuning (NT), low-tuning

(LT), and high-tuning (HT). Bright regions in the NT, LT, and

HT images represent infarction (or blood), akinetic, and

contracting tissues, respectively. An anatomy (ANAT) image

of the heart can also be constructed by adding the LT and HT

images as described in 0 to show the anatomical structure of

the heart (both contracting and non-contracting myocardium)

with no signal from blood. Figure 1 shows an example of

acquired C-SENC images.

Figure 1. C-SENC Images of a patient suffering myocardial infarction: a)

NT, b) LT, c) HT

A

a) b) c)



UniCSE 1 (1), 36 -40, 2010

37

Table I shows the signal intensities in C-SENC images for

the blood, background, and the different tissue types.

B. Otsu Thresholding

Thresholding is a well-known technique for imagesegmentation that tries to identify and extract a target from itsbackground on the basis of the distribution of gray levels in

image objects. Otsu’s method chooses the optimal thresholds by maximizing the between-class variance with an exhaustivesearch [7].

Reference [8] concluded that Otsu’s method was one of the

best threshold selection methods for general real world images.Otsu’s method uses an exhaustive search to evaluate the criterion for maximizing the between-class variance.

An image is a 2D grayscale intensity function, and containsN pixels with gray levels from 1 to L. The number of pixelswith gray level i is denoted ni, giving a probability of gray leveli in an image of

⁄

In the case of bi-level thresholding of an image, the pixelsare divided into two classes, C1 with gray levels [1, …, t] and C2 with gray levels [t+1, …, L]. Then, the gray levelprobability distributions for the two classes are

⁄ ⁄ , and

⁄ ⁄ ⁄

where ∑

and ∑

Also, the means for classes C1 and C2 are

∑ ⁄ and (4)

∑ ⁄

Let µ T be the mean intensity for the whole image. It is easyto show that

Otsu defined the between-class variance [7] of thethresholded image as

For bi-level thresholding, Otsu verified that the optimalthreshold t* is chosen so that the between-class variance ismaximized; that is,

{} where

C. Theo Pavlidis’ Tracing Algorithm

This algorithm is one of the relatively recent contourtracing algorithms. It was proposed by Theo Pavlidis in1982 [9]. In this algorithm, the contour is traced in acounterclockwise direction. This only affects the relativedirection of movements while tracing the contour.

Given a digital binary image (a grid of black pixels, on a

background of white pixels); a black pixel is located anddeclared as a "start" pixel. Locating a "start" pixel can be donein a number of ways; one of which is done by starting at thebottom left corner of the grid, scanning each column of pixelsfrom the bottom going upwards until a black pixel isencountered.

Any black boundary pixel can be chosen to bethe start pixel as long as its left adjacent pixel is white.Throughout the algorithm, the pixels of interest at any time arethe three pixels: P1, P2 and P3 shown in Figure 2.

Figure 2. The three pixels of interest when scanning the pixel P0. P2 is the

pixel in front of P0, P1 is the pixel adjacent to P2 from the left and P3 is the

right adjacent pixel to P2.

The steps of the algorithm can be summarized as follows:

When the current boundary pixel is P0, pixel P1 ischecked. If P1 is black, then declare P1 to be thecurrent boundary pixel and move one step forward

TABLE I

TISSUE TYPES OCCUPY DIFFERENT AND THEIR CHARACTERISTIC SIGNAL

INTENSITIES IN THE USED IMAGES

Type ANAT NT LT HT

Blood Low High Low Low

Infarcted High High High Low

Functional High Medium low High

Background Low Low Low Low

Signal intensities of different tissues in the No-Tuning (NT),

Anatomy (ANAT), and High-Tuning (HT) C-SENC images. High = signal intensities greater than 0.8 in normalized

images; and Low = signal intensities less than 0.2 in

normalized images.



UniCSE 1 (1), 36 -40, 2010

38

followed by one step to the left to reachP1. Whereas if P1 is white proceed to check P2.

If P2 is black, then P2 is declared to be the currentboundary pixel. If P1 and P2 are white, check P3.

If P3 is black, then P3 is declared to be the currentboundary pixel.

If all the three pixels (P1, P2, P3) in front of the

current boundary pixel (P0) are white, then, rotate(while standing on the current boundary pixel) 90degrees clockwise to face a new set of pixels infront of P0.

Repeat the previous four steps until no black pixelsare found or the current boundary pixel is the startone.

D. Subtractive Clustering

Subtractive clustering [10] uses the positions of the given

data points to calculate a density function instead of

calculating the density function at every possible position in

the data space, thus reducing the number of calculations

significantly. It uses data points as the candidates for clustercenters. However, the actual cluster centers are not necessarily

located at one of the data points, but in most cases it is a good

approximation.

Since each data point is a candidate for cluster centers, a

density measure at data point is defined as

( ‖ ‖ ⁄ )

where is a positive constant representing a neighborhood

radius. Hence, a data point will have a high density value if it

has many neighboring data points.The first cluster center is chosen as the point having the

largest density value . Next, the density measure of each

data point is revised as follows:

( ‖ ‖ ⁄ )

where is a positive constant which defines a

neighborhood that has measurable reductions in density

measure. Therefore, the data points near the first cluster centerwill have significantly reduced density measure.

After revising the density function, the next cluster center isselected as the point having the greatest density value. Thisprocess continues until a sufficient number of clusters isattainted.

III. MATERIALS AND METHODS

A. The Proposed Method

Figure 3 shows a sequence diagram of the proposedmethod, which consists of three main stages. The first stage of the proposed technique aims to detect and extract the leftventricle. This is done by the successive application of threesub-steps: thresholding, morphological opening and boundarydetection.

Threshold grey-level intensity value is calculated usingOtsu’s thresholding. This value is applied to the ANAT image

to binarize it and to differentiate pixels representing tissue fromthose representing background. Then, morphological openingis applied on the thresholded image. This aims to remove morenoise pixels and to improve the myocardium boundaries so thatthey can be traced successfully in the next step. This step isapplied for two successive times; the first time removes thesmall holes in the tissue while the second time does the samefor the small holes in the background region.

Figure 3. Stages of the Proposed Method

Finally, boundary detection is applied to select the leftventricular borders. This is done by using Theo Pavlidis’ tracing algorithm [9]. All regions are traced, and then the mostcentralized region is selected to be the left ventricle as shownin Figure 4. This results in eliminating all regions that representadditional tissue areas.



UniCSE 1 (1), 36 -40, 2010

39

Figure 4. a) Morphologically opened image, b) image with boundary tracedregions, c) Image showing the most centralized region that represents the left

ventricle

In the second stage, the ANAT image of the heart isconstructed by adding the LT and HT images as described in 0.Then, NT and ANAT Images are masked by the myocardiumboundariesextracted in the second stage to exclude all pixelsoutside the left ventricle. This restricts the clustering to onlytwo pixel types; functional and infarcted tissue. This is appliedby ANDING the binary image that contains the pixels of theleft ventricle only with the original ANAT and NT images,respectively. This results in two images (ANAT and NT) thatcontain only pixels corresponding to the left ventricular tissue.

The third stage of the proposed method consists of further

clustering the myocardium through the application of subtractive clustering algorithm. In this stage, myocardium isclassified into two major clusters: contracting and non-contracting tissue, from which we can extract the infarctedregions. Subtractive clustering detects the number of clustersand the central pixel of each one. This means preserving themethod from failure when it is applied to real cases of fullyfunctional heart, i.e. there is no infarcted tissue areas

B. Validation

A consultant cardiologist helped in the validation of theproposed method by marking the infarcted tissue regions in thereal C-SENC MRI images. Infarcted regions were marked threetimes for each set. Infarctions extracted from the clusteredimages were compared with those extracted from the markedimages on a pixel-by-pixel basis and statistical results werecalculated. Figure 5 shows a colored scatter plot of the NT andANAT images.

Figure 5. Fig. 5. Scatter plot of normalized NT and ANAT Images. The red

scatters represent the infarcted pixels while the blue ones represent thefunction tissue.

IV. RESULTS

Figure 6 shows real short-axis C-SENC images and theresulting infarcted regions colored in red. The resulting imagescorrectly identified the pixels marked as infarct with sensitivityof 90.19 – 98.38 % and specificity of 77.13 – 93.77 %.Quantitative performance and visual comparisons demonstratethe robustness of the proposed technique for identifyingdifferent tissue types.

Figure 6. Representative C-SENC MRI images from a patient with

myocardial infarction. The resulting infarcted regions are colored in red. Thefirst column (d, g and j) shows a sample real images case. and the infarction

map for the proposed technique. The second column (e, h and k) shows the

markings of the consultant cardiologist. The third column (f, i and l) shows theresulting markings of the proposed method. Accuracy = 97.22%, Precision =

84.01%.

V. DISCUSSION AND CONCLUSION

The proposed technique was found to be robust indetermining the existent infarction. Infarcted myocardium isspecified only if there are enough data points in the infarct

cluster during the application of the technique. For all theanalyzed images, the technique correctly determines infarctionexistence.

One advantage of the resulting clustered images is theexcellent removal of the blood due to the black-blood propertyin the LT and HT images. Thus, the resulting clustered imageswould allow for accurate measurement of the infarct size.Another advantage is that the proposed method can detect thenumber of clusters in the given images. This allows forsuccessful detection of infarction absence.

a) b) c)

a) b) c)

d) e) f)

g) h) i)

j) k)



UniCSE 1 (1), 36 -40, 2010

40

Tissue of the myocardium detected as infarcted tissueindeed includes two clusters; infarct and hibernated (tissue thatis not-contracting while still viable). It is clear that we are ableto differentiate between contracting and non-contractingregions in the heart tissue, but still cannot differentiate betweenthe hibernating and the infracted tissue as both of them are non-contracting tissues. Further clustering may be applied todifferentiate between these two regions.

In conclusion, a new method is proposed for identifyingdifferent heart tissues from C-SENC functional and viabilityimages. The technique is based on the consecutive applicationof the Otsu’s thresholding and subtractive clusteringtechniques. The method is successfully applied to transversecross sectional images of the heart. It gives very good resultsregarding sensitivity and specificity although this type of images is known to suffer from high noise levels.

REFERENCES

[1] Ibrahim EH, Stuber M, Kraitchman DL, Weiss RG, Osman NF.Combined functional and viability cardiac MR imaging in a singlebreathhold. Magn Reson Med 2007; 58:843 – 849.

[2] Osman NF, Sampath S, Atalar E, Prince JL. Imaging longitudinal

cardiac strain on short-axis images using strain-encoding MRI. MagnReson Med 2001; 46:324 – 334.

[3] Wu KC, Lima JA. Noninvasive imaging of myocardial viability: currenttechniques and future developments. Circ Res 2003; 93:1146 – 1158.

[4] Watzinger N, Saeed M, Wendland MF, Akbari H, Lund G, Higgins CB.Myocardial viability: magnetic resonance assessment of functionalreserve and tissue characterization. J Cardiovasc Magn Reson 2001;3:195 – 208.

[5] Ibrahim, E.-S.H., Weiss, R.G., Stuber, M., Spooner, A.E., Osman, N.F.:Identification of Different Heart Tissues from MRI C-SENC ImagesUsing an unsupervised Multi-Stage Fuzzy Clustering Technique. J.Magn. Reson. Imaging 28(2), 519 – 526 (2008).

[6] Motaal, A.G., El Gayar, N., Osman, N.F.: Automated Cardiac-TissueIdentification in Composite Strain-Encoded (C-SENC) Images UsingFuzzy C-means and Bayesian Classifier. iCBBE, Chengdu, China, 2010.

[7] N. Otsu, A threshold selection method from gray-level histogram, IEEETransactions on System Man Cybernetics, Vol. SMC-9, No. 1, 1979, pp.62-66.

[8] Neamat El Gayar, Friedhelm Schwenker, and Gunther Palm. A Study of the Robustness of KNN Classifiers Trained Using Soft Labels. ANNPR 2006, LNAI 4087, pp. 67 – 80, 2006.

[9] T. Pavlidis, Algorithms for Graphics and Image Processing. Berlin,Germany: Springer Verlag, 1982, ch. 7.

[10] Jang, J.-S. R., Sun, C.-T., Mizutani, E., Neuro-Fuzzy and SoftComputing – A Computational Approach to Learning and MachineIntelligence, Prentice Hall, 1997.

Detection of Cardiac Infarction in MRI C-SENC Images

Documents

Transcript of Detection of Cardiac Infarction in MRI C-SENC Images